<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: SK Faizanuddin</title>
    <description>The latest articles on DEV Community by SK Faizanuddin (@skfaizan786).</description>
    <link>https://dev.to/skfaizan786</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F3800234%2Ff30680ba-5abd-4a60-b842-7e98fffcee8c.jpeg</url>
      <title>DEV Community: SK Faizanuddin</title>
      <link>https://dev.to/skfaizan786</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/skfaizan786"/>
    <language>en</language>
    <item>
      <title>Why LLM Code Reviewers Fail at Architecture—And How We Fixed It with a 9ms FalkorDB Knowledge Graph</title>
      <dc:creator>SK Faizanuddin</dc:creator>
      <pubDate>Sun, 18 Oct 2026 12:47:25 +0000</pubDate>
      <link>https://dev.to/skfaizan786/why-llm-code-reviewers-fail-at-architecture-and-how-we-fixed-it-with-a-9ms-falkordb-knowledge-graph-144m</link>
      <guid>https://dev.to/skfaizan786/why-llm-code-reviewers-fail-at-architecture-and-how-we-fixed-it-with-a-9ms-falkordb-knowledge-graph-144m</guid>
      <description>&lt;h2&gt;
  
  
  The Friday Afternoon Outage Story
&lt;/h2&gt;

&lt;p&gt;It was 4:15 PM on a Friday when an engineer on the identity team pushed a seemingly harmless PR to &lt;code&gt;auth-service&lt;/code&gt;:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight diff"&gt;&lt;code&gt;&lt;span class="gd"&gt;--- a/schemas/user.proto
&lt;/span&gt;&lt;span class="gi"&gt;+++ b/schemas/user.proto
&lt;/span&gt;&lt;span class="p"&gt;@@ -12,3 +12,3 @@&lt;/span&gt; message UserProfile {
&lt;span class="gd"&gt;-  string user_id = 1; // UUIDv4 RFC 4122
&lt;/span&gt;&lt;span class="gi"&gt;+  int64 user_id = 1; // Optimized sequential 64-bit ID
&lt;/span&gt; }
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The pull request was titled: &lt;em&gt;"Optimize user lookup indexing: switch UUIDv4 strings to sequential 64-bit integers"&lt;/em&gt;.&lt;/p&gt;

&lt;p&gt;Every CI check passed green:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;TypeScript typechecks passed inside &lt;code&gt;auth-service&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;Protobuf generated clean stubs.&lt;/li&gt;
&lt;li&gt;Unit tests executed cleanly with zero syntax failures.&lt;/li&gt;
&lt;li&gt;An AI code-review bot left a cheerful comment: &lt;em&gt;"Looks clean! Sequential IDs reduce B-tree index fragmentation by up to 25%."&lt;/em&gt;
&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Twenty minutes after merging to staging, the payments gateway collapsed.&lt;/p&gt;

&lt;p&gt;What the CI runner and the AI reviewer couldn't see was that two years ago, the architecture board ratified &lt;strong&gt;ADR-001&lt;/strong&gt;: &lt;em&gt;"Canonical User Identification Standard"&lt;/em&gt;, strictly mandating UUIDv4 strings across all internal APIs to prevent distributed collision attacks and enable multi-tenant data exports.&lt;/p&gt;

&lt;p&gt;Even worse, three hops away in the microservice topology:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;code&gt;auth-service&lt;/code&gt; exposes &lt;code&gt;/v1/users&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;checkout-service&lt;/code&gt; consumes &lt;code&gt;/v1/users&lt;/code&gt; to authorize customer carts.&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;billing-service&lt;/code&gt; and &lt;code&gt;ledger-service&lt;/code&gt; deserialize &lt;code&gt;user_id&lt;/code&gt; into strict UUID data stores.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;By changing one field in an upstream schema, the PR silently bricked three downstream services, disrupted checkout flows, and dragged two engineering teams into an emergency triage incident.&lt;/p&gt;

&lt;p&gt;Linters only check syntax. AI reviewers only evaluate isolated files. Neither has a clue how your company actually connects.&lt;/p&gt;

&lt;p&gt;That is why I built &lt;strong&gt;ArchGuard&lt;/strong&gt; for the &lt;strong&gt;WeMakeDevs x FalkorDB Graph Hacks Hackathon&lt;/strong&gt; under the &lt;strong&gt;Company Brain &amp;amp; Episodic Memory Track&lt;/strong&gt;.&lt;/p&gt;




&lt;h2&gt;
  
  
  Why Vector RAG &amp;amp; Giant Context Windows Fail at Architecture
&lt;/h2&gt;

&lt;p&gt;The natural modern reaction to this problem is: &lt;em&gt;"Just throw all the ADRs into an LLM context window or build a Vector RAG pipeline."&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;When I started designing ArchGuard, I tested that exact approach. It failed immediately for three fundamental reasons:&lt;/p&gt;

&lt;h3&gt;
  
  
  1. Vector Similarity Searches for Topics, Not Invariant Rules
&lt;/h3&gt;

&lt;p&gt;Vector embeddings measure semantic closeness. If a pull request modifies &lt;code&gt;int64 user_id&lt;/code&gt;, vector search queries like &lt;code&gt;"user_id data type policy"&lt;/code&gt; might return a blog post about database migration or an RFC on user profiles. But architectural governance is &lt;strong&gt;binary and relational&lt;/strong&gt;:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Does &lt;code&gt;ADR-001&lt;/code&gt; govern &lt;code&gt;user_id&lt;/code&gt;? &lt;strong&gt;Yes.&lt;/strong&gt;
&lt;/li&gt;
&lt;li&gt;Does &lt;code&gt;int64&lt;/code&gt; match &lt;code&gt;ADR-001&lt;/code&gt;'s expected format &lt;code&gt;UUIDv4&lt;/code&gt;? &lt;strong&gt;No.&lt;/strong&gt;
&lt;/li&gt;
&lt;li&gt;An LLM guesses based on probabilities. A graph proves it with mathematical certainty.&lt;/li&gt;
&lt;/ul&gt;

&lt;h3&gt;
  
  
  2. Multi-Hop Transitive Blast Radius Is Invisible to RAG
&lt;/h3&gt;

&lt;p&gt;In a microservice mesh, dependencies are directed paths:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight cypher"&gt;&lt;code&gt;&lt;span class="ss"&gt;(&lt;/span&gt;&lt;span class="n"&gt;PR&lt;/span&gt;&lt;span class="ss"&gt;)&lt;/span&gt; &lt;span class="err"&gt;➔&lt;/span&gt; &lt;span class="ss"&gt;(&lt;/span&gt;&lt;span class="nc"&gt;:SchemaField&lt;/span&gt; &lt;span class="ss"&gt;{&lt;/span&gt;&lt;span class="py"&gt;name:&lt;/span&gt; &lt;span class="s2"&gt;"user_id"&lt;/span&gt;&lt;span class="ss"&gt;})&lt;/span&gt;
          &lt;span class="err"&gt;▲&lt;/span&gt;
          &lt;span class="err"&gt;│&lt;/span&gt; &lt;span class="ss"&gt;[&lt;/span&gt;&lt;span class="nc"&gt;:DEFINED_IN&lt;/span&gt;&lt;span class="ss"&gt;]&lt;/span&gt;
     &lt;span class="ss"&gt;(&lt;/span&gt;&lt;span class="nc"&gt;:Service&lt;/span&gt; &lt;span class="ss"&gt;{&lt;/span&gt;&lt;span class="py"&gt;name:&lt;/span&gt; &lt;span class="s2"&gt;"auth-service"&lt;/span&gt;&lt;span class="ss"&gt;})&lt;/span&gt;
          &lt;span class="err"&gt;▲&lt;/span&gt;
          &lt;span class="err"&gt;│&lt;/span&gt; &lt;span class="ss"&gt;[&lt;/span&gt;&lt;span class="nc"&gt;:CALLS&lt;/span&gt;&lt;span class="ss"&gt;]&lt;/span&gt;
     &lt;span class="ss"&gt;(&lt;/span&gt;&lt;span class="nc"&gt;:Service&lt;/span&gt; &lt;span class="ss"&gt;{&lt;/span&gt;&lt;span class="py"&gt;name:&lt;/span&gt; &lt;span class="s2"&gt;"checkout-service"&lt;/span&gt;&lt;span class="ss"&gt;})&lt;/span&gt;
          &lt;span class="err"&gt;▲&lt;/span&gt;
          &lt;span class="err"&gt;│&lt;/span&gt; &lt;span class="ss"&gt;[&lt;/span&gt;&lt;span class="nc"&gt;:DEPENDS_ON&lt;/span&gt;&lt;span class="ss"&gt;]&lt;/span&gt;
     &lt;span class="ss"&gt;(&lt;/span&gt;&lt;span class="nc"&gt;:Service&lt;/span&gt; &lt;span class="ss"&gt;{&lt;/span&gt;&lt;span class="py"&gt;name:&lt;/span&gt; &lt;span class="s2"&gt;"billing-service"&lt;/span&gt;&lt;span class="ss"&gt;})&lt;/span&gt;
          &lt;span class="err"&gt;│&lt;/span&gt;
          &lt;span class="err"&gt;▼&lt;/span&gt; &lt;span class="ss"&gt;[&lt;/span&gt;&lt;span class="nc"&gt;:OWNED_BY&lt;/span&gt;&lt;span class="ss"&gt;]&lt;/span&gt;
     &lt;span class="ss"&gt;(&lt;/span&gt;&lt;span class="nc"&gt;:Team&lt;/span&gt; &lt;span class="ss"&gt;{&lt;/span&gt;&lt;span class="py"&gt;name:&lt;/span&gt; &lt;span class="s2"&gt;"Billing Team"&lt;/span&gt;&lt;span class="ss"&gt;})&lt;/span&gt; &lt;span class="err"&gt;➔&lt;/span&gt; &lt;span class="ss"&gt;(&lt;/span&gt;&lt;span class="err"&gt;@&lt;/span&gt;&lt;span class="py"&gt;elena_fin:&lt;/span&gt; &lt;span class="n"&gt;Person&lt;/span&gt;&lt;span class="ss"&gt;)&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;No vector database can follow that 4-hop chain in a single step to tell you: &lt;em&gt;"Changing this protobuf field breaks Elena's billing pipeline 3 services away."&lt;/em&gt;&lt;/p&gt;

&lt;h3&gt;
  
  
  3. Context Window Bloat Destroys Attention
&lt;/h3&gt;

&lt;p&gt;An enterprise with 60 microservices, 45 ADRs, and hundreds of OpenAPI specifications generates megabytes of technical documentation. Shoving all of that into a prompt on every pull request is expensive, slow (15–30 seconds), and suffers from catastrophic attention degradation where the model ignores critical bullet points buried in the middle.&lt;/p&gt;




&lt;h2&gt;
  
  
  Enter ArchGuard: The Company Brain Architecture
&lt;/h2&gt;

&lt;p&gt;ArchGuard flips the paradigm: &lt;strong&gt;Deterministic Detection, LLM-Explained.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;Instead of asking an LLM to guess whether an architectural decision was violated, ArchGuard compiles company architecture into a high-performance &lt;strong&gt;FalkorDB Knowledge Graph&lt;/strong&gt; (&lt;code&gt;archguard_brain&lt;/code&gt;). &lt;/p&gt;

&lt;p&gt;FalkorDB handles deterministic graph traversals in &lt;strong&gt;under 15 milliseconds&lt;/strong&gt; per Cypher query (with the full multi-agent graph pipeline executing in ~23ms–108ms, compared to 4.0s for LLM synthesis). The LLM's only job is to translate the raw input diff into AST parameters and synthesize a human-friendly remediation plan from historical precedents.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fzzdduixap7fywyjrapdf.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fzzdduixap7fywyjrapdf.png" alt="ArchGuard Constellation Canvas" width="800" height="450"&gt;&lt;/a&gt;&lt;br&gt;
&lt;em&gt;Figure 1: ArchGuard Command Center showing real-time constellation topology, isolated collision path (red), and multi-hop blast radius.&lt;/em&gt;&lt;/p&gt;
&lt;h3&gt;
  
  
  The FinCore Domain Knowledge Graph
&lt;/h3&gt;

&lt;p&gt;The system ships with a production-grade fintech topology (&lt;strong&gt;FinCore Financial Systems&lt;/strong&gt;) modeled directly in FalkorDB:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;                            FINCORE TOPOLOGY IN FALKORDB

     🟡 [ ADR-001: UUIDv4 ]             🟡 [ ADR-002: Async Queue ]
              │                                      │
       [:ENFORCES]                              [:GOVERNS]
              ▼                                      ▼
    🟡 (user_id: SchemaField)              🔵 (/v1/webhooks/payment)
              │                                      │
        [:DEFINED_IN]                           [:EXPOSED_BY]
              ▼                                      ▼
      🔵 [ auth-service ]                    🔵 [ payment-gateway ]
              ▲                                      │
              │ [:CALLS]                             │ [:DEPENDS_ON]
      🔵 [ checkout-service ] ◄──────────────────────┘
              │
         [:OWNED_BY]
              ▼
      🟢 { Team: Checkout } ◄───[:LEADS]─── 🟢 (@dave_eng: Person)
              ▲
              │ [:APPLIES_TO]
      🧠 [ INC-2023-42: Legacy Workaround ]
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The baseline graph contains &lt;strong&gt;28 pristine nodes and 35 directed relationships&lt;/strong&gt;:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;🟡 &lt;strong&gt;Golden Rules &amp;amp; Contracts:&lt;/strong&gt; 4 &lt;code&gt;:ADR&lt;/code&gt; nodes and 3 &lt;code&gt;:SchemaField&lt;/code&gt; nodes.&lt;/li&gt;
&lt;li&gt;🟢 &lt;strong&gt;Organizational Context:&lt;/strong&gt; 4 &lt;code&gt;:Team&lt;/code&gt; nodes and 5 &lt;code&gt;:Person&lt;/code&gt; nodes (Tech Leads).&lt;/li&gt;
&lt;li&gt;🔵 &lt;strong&gt;Distributed Infrastructure:&lt;/strong&gt; 5 &lt;code&gt;:Service&lt;/code&gt; nodes and 5 &lt;code&gt;:Endpoint&lt;/code&gt; nodes.&lt;/li&gt;
&lt;li&gt;🔴 &lt;strong&gt;Transient PR State:&lt;/strong&gt; Dynamic &lt;code&gt;:Proposal&lt;/code&gt; nodes created on-the-fly.&lt;/li&gt;
&lt;li&gt;🧠 &lt;strong&gt;Cross-Session Episodic Memory:&lt;/strong&gt; 2 &lt;code&gt;:ArchitecturalEpisode&lt;/code&gt; nodes documenting historical incident precedents and approved trade-offs.&lt;/li&gt;
&lt;/ul&gt;




&lt;h2&gt;
  
  
  How 3 Autonomous Agents Share State Through FalkorDB
&lt;/h2&gt;

&lt;p&gt;Most multi-agent frameworks concatenate prompts: Agent A writes a huge JSON string, pastes it into Agent B's prompt, which appends another blob for Agent C. This causes compounding hallucinations and context explosion.&lt;/p&gt;

&lt;p&gt;In ArchGuard, our 3 agents &lt;strong&gt;never talk directly to each other&lt;/strong&gt;. They communicate strictly by reading and mutating shared subgraphs inside FalkorDB.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;                              AGENT GRAPH PIPELINE

   Git Diff / OpenAPI Spec
              │
              ▼
    ┌──────────────────┐
    │   AGENT 1        │  1. Ingests raw diff
    │  Spec Extractor  │  2. Identifies target field/endpoint
    └─────────┬────────┘  3. Executes Cypher MERGE
              │
              ▼
      ┌───────────────┐   MERGE (:Proposal {id: $id, status: 'PARSED'})
      │   FalkorDB    │   CREATE (p)-[:TARGETS_FIELD]-&amp;gt;(target)
      │  Graph State  │
      └───────┬───────┘
              │
              ▼ (Passes ONLY proposalId)
    ┌──────────────────┐
    │   AGENT 2        │  1. Runs deterministic Cypher traversal (&amp;lt;15ms)
    │  Graph Auditor   │  2. Discovers ADR collisions &amp;amp; broken services
    └─────────┬────────┘  3. Writes findings directly to graph
              │
              ▼
      ┌───────────────┐   CREATE (p)-[:VIOLATES_ADR]-&amp;gt;(adr)
      │   FalkorDB    │   CREATE (p)-[:BREAKS_CONSUMER]-&amp;gt;(service)
      │  Graph State  │   SET p.status = 'CHECKED'
      └───────┬───────┘
              │
              ▼ (Passes ONLY proposalId)
    ┌──────────────────┐
    │   AGENT 3        │  1. Queries (:Proposal)-[:VIOLATES_ADR]-&amp;gt;(:ADR)&amp;lt;-[:APPLIES_TO]-(:Episode)
    │  Advisor Memory  │  2. Recalls past incident compromise &amp;amp; approved workaround
    └─────────┬────────┘  3. Synthesizes compliant migration plan &amp;amp; replacement diff
              │
              ▼
      ┌───────────────┐   SET p.status = 'RESOLVED'
      │   FalkorDB    │   (Optional: Commits new :ArchitecturalEpisode)
      └───────────────┘
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Notice how clean the orchestrator implementation is in &lt;code&gt;src/agents/orchestrator.ts&lt;/code&gt;:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight typescript"&gt;&lt;code&gt;&lt;span class="c1"&gt;// Agent 1: Parses diff and creates :Proposal node in FalkorDB&lt;/span&gt;
&lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="nx"&gt;extractorOutput&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="k"&gt;await&lt;/span&gt; &lt;span class="nf"&gt;runExtractorAgent&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nx"&gt;rawInput&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nx"&gt;title&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nx"&gt;options&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nx"&gt;tenant&lt;/span&gt;&lt;span class="p"&gt;);&lt;/span&gt;
&lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="nx"&gt;proposalId&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nx"&gt;extractorOutput&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nx"&gt;proposalId&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;

&lt;span class="c1"&gt;// Agent 2: Receives ONLY the proposalId; reads and writes graph state&lt;/span&gt;
&lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="nx"&gt;auditorOutput&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="k"&gt;await&lt;/span&gt; &lt;span class="nf"&gt;runAuditorAgent&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nx"&gt;proposalId&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nx"&gt;tenant&lt;/span&gt;&lt;span class="p"&gt;);&lt;/span&gt;

&lt;span class="c1"&gt;// Agent 3: Receives ONLY the proposalId; recalls episodic memory from graph&lt;/span&gt;
&lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="nx"&gt;advisorOutput&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="k"&gt;await&lt;/span&gt; &lt;span class="nf"&gt;runAdvisorAgent&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nx"&gt;proposalId&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nx"&gt;options&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nx"&gt;tenant&lt;/span&gt;&lt;span class="p"&gt;);&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Neither Agent 2 nor Agent 3 needs the raw git diff in memory. They query FalkorDB for the &lt;code&gt;:Proposal&lt;/code&gt; node, inspect its relationships, and write their verdicts directly to the graph.&lt;/p&gt;




&lt;h2&gt;
  
  
  The Real Cypher Queries Doing the Heavy Lifting
&lt;/h2&gt;

&lt;p&gt;Here are the actual Cypher queries executed inside &lt;code&gt;src/engine/graphAuditor.ts&lt;/code&gt; that make ArchGuard lightning fast:&lt;/p&gt;

&lt;h3&gt;
  
  
  1. Deterministic ADR Invariant Verification
&lt;/h3&gt;

&lt;p&gt;When a proposed change targets a schema field or endpoint, Agent 2 checks if any active ADR governs that target with a conflicting invariant:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight cypher"&gt;&lt;code&gt;&lt;span class="k"&gt;MATCH&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="ss"&gt;(&lt;/span&gt;&lt;span class="py"&gt;p:&lt;/span&gt;&lt;span class="n"&gt;Proposal&lt;/span&gt; &lt;span class="ss"&gt;{&lt;/span&gt;&lt;span class="py"&gt;id:&lt;/span&gt; &lt;span class="n"&gt;$proposalId&lt;/span&gt;&lt;span class="ss"&gt;})&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="ss"&gt;[&lt;/span&gt;&lt;span class="py"&gt;t:&lt;/span&gt;&lt;span class="n"&gt;TARGETS_FIELD&lt;/span&gt;&lt;span class="ss"&gt;]&lt;/span&gt;&lt;span class="o"&gt;-&amp;gt;&lt;/span&gt;&lt;span class="ss"&gt;(&lt;/span&gt;&lt;span class="n"&gt;target&lt;/span&gt;&lt;span class="ss"&gt;)&lt;/span&gt;&lt;span class="o"&gt;&amp;lt;-&lt;/span&gt;&lt;span class="ss"&gt;[&lt;/span&gt;&lt;span class="nc"&gt;:GOVERNS&lt;/span&gt;&lt;span class="ss"&gt;]&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="ss"&gt;(&lt;/span&gt;&lt;span class="py"&gt;adr:&lt;/span&gt;&lt;span class="n"&gt;ADR&lt;/span&gt;&lt;span class="ss"&gt;)&lt;/span&gt;
&lt;span class="k"&gt;WHERE&lt;/span&gt; &lt;span class="n"&gt;adr.target_property&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;t.changed_property&lt;/span&gt; 
  &lt;span class="ow"&gt;AND&lt;/span&gt; &lt;span class="n"&gt;adr.expected_value&lt;/span&gt; &lt;span class="o"&gt;&amp;lt;&amp;gt;&lt;/span&gt; &lt;span class="n"&gt;t.proposed_value&lt;/span&gt;
&lt;span class="k"&gt;OPTIONAL&lt;/span&gt; &lt;span class="k"&gt;MATCH&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="ss"&gt;(&lt;/span&gt;&lt;span class="n"&gt;adr&lt;/span&gt;&lt;span class="ss"&gt;)&lt;/span&gt;&lt;span class="o"&gt;&amp;lt;-&lt;/span&gt;&lt;span class="ss"&gt;[&lt;/span&gt;&lt;span class="nc"&gt;:AUTHORED&lt;/span&gt;&lt;span class="ss"&gt;]&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="ss"&gt;(&lt;/span&gt;&lt;span class="py"&gt;author:&lt;/span&gt;&lt;span class="n"&gt;Person&lt;/span&gt;&lt;span class="ss"&gt;)&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="ss"&gt;[&lt;/span&gt;&lt;span class="nc"&gt;:MEMBER_OF&lt;/span&gt;&lt;span class="ss"&gt;]&lt;/span&gt;&lt;span class="o"&gt;-&amp;gt;&lt;/span&gt;&lt;span class="ss"&gt;(&lt;/span&gt;&lt;span class="py"&gt;authorTeam:&lt;/span&gt;&lt;span class="n"&gt;Team&lt;/span&gt;&lt;span class="ss"&gt;)&lt;/span&gt;
&lt;span class="k"&gt;RETURN&lt;/span&gt; 
  &lt;span class="n"&gt;adr.code&lt;/span&gt; &lt;span class="k"&gt;AS&lt;/span&gt; &lt;span class="n"&gt;adrCode&lt;/span&gt;&lt;span class="ss"&gt;,&lt;/span&gt;
  &lt;span class="n"&gt;adr.title&lt;/span&gt; &lt;span class="k"&gt;AS&lt;/span&gt; &lt;span class="n"&gt;adrTitle&lt;/span&gt;&lt;span class="ss"&gt;,&lt;/span&gt;
  &lt;span class="n"&gt;adr.expected_value&lt;/span&gt; &lt;span class="k"&gt;AS&lt;/span&gt; &lt;span class="n"&gt;expectedValue&lt;/span&gt;&lt;span class="ss"&gt;,&lt;/span&gt;
  &lt;span class="n"&gt;t.proposed_value&lt;/span&gt; &lt;span class="k"&gt;AS&lt;/span&gt; &lt;span class="n"&gt;proposedValue&lt;/span&gt;&lt;span class="ss"&gt;,&lt;/span&gt;
  &lt;span class="n"&gt;author.name&lt;/span&gt; &lt;span class="k"&gt;AS&lt;/span&gt; &lt;span class="n"&gt;authorName&lt;/span&gt;&lt;span class="ss"&gt;,&lt;/span&gt;
  &lt;span class="n"&gt;authorTeam.name&lt;/span&gt; &lt;span class="k"&gt;AS&lt;/span&gt; &lt;span class="n"&gt;authorTeam&lt;/span&gt;&lt;span class="ss"&gt;,&lt;/span&gt;
  &lt;span class="n"&gt;authorTeam.slack_channel&lt;/span&gt; &lt;span class="k"&gt;AS&lt;/span&gt; &lt;span class="n"&gt;slackChannel&lt;/span&gt;&lt;span class="ss"&gt;;&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;If a match is found, Agent 2 immediately executes:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight cypher"&gt;&lt;code&gt;&lt;span class="k"&gt;MATCH&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="ss"&gt;(&lt;/span&gt;&lt;span class="py"&gt;p:&lt;/span&gt;&lt;span class="n"&gt;Proposal&lt;/span&gt; &lt;span class="ss"&gt;{&lt;/span&gt;&lt;span class="py"&gt;id:&lt;/span&gt; &lt;span class="n"&gt;$proposalId&lt;/span&gt;&lt;span class="ss"&gt;}),&lt;/span&gt; &lt;span class="ss"&gt;(&lt;/span&gt;&lt;span class="py"&gt;a:&lt;/span&gt;&lt;span class="n"&gt;ADR&lt;/span&gt; &lt;span class="ss"&gt;{&lt;/span&gt;&lt;span class="py"&gt;code:&lt;/span&gt; &lt;span class="n"&gt;$adrCode&lt;/span&gt;&lt;span class="ss"&gt;})&lt;/span&gt;
&lt;span class="k"&gt;MERGE&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="ss"&gt;(&lt;/span&gt;&lt;span class="n"&gt;p&lt;/span&gt;&lt;span class="ss"&gt;)&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="ss"&gt;[&lt;/span&gt;&lt;span class="py"&gt;r:&lt;/span&gt;&lt;span class="n"&gt;VIOLATES_ADR&lt;/span&gt; &lt;span class="ss"&gt;{&lt;/span&gt;&lt;span class="py"&gt;reason:&lt;/span&gt; &lt;span class="n"&gt;$reason&lt;/span&gt;&lt;span class="ss"&gt;,&lt;/span&gt; &lt;span class="py"&gt;detected_at:&lt;/span&gt; &lt;span class="n"&gt;datetime&lt;/span&gt;&lt;span class="ss"&gt;()}]&lt;/span&gt;&lt;span class="o"&gt;-&amp;gt;&lt;/span&gt;&lt;span class="ss"&gt;(&lt;/span&gt;&lt;span class="n"&gt;a&lt;/span&gt;&lt;span class="ss"&gt;)&lt;/span&gt;
&lt;span class="k"&gt;SET&lt;/span&gt; &lt;span class="n"&gt;p.status&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="s1"&gt;'CHECKED'&lt;/span&gt;&lt;span class="ss"&gt;;&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;h3&gt;
  
  
  2. Multi-Hop Blast Radius Traversal
&lt;/h3&gt;

&lt;p&gt;Next, Agent 2 determines which downstream microservices and tech leads will be impacted:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight cypher"&gt;&lt;code&gt;&lt;span class="k"&gt;MATCH&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="ss"&gt;(&lt;/span&gt;&lt;span class="py"&gt;p:&lt;/span&gt;&lt;span class="n"&gt;Proposal&lt;/span&gt; &lt;span class="ss"&gt;{&lt;/span&gt;&lt;span class="py"&gt;id:&lt;/span&gt; &lt;span class="n"&gt;$proposalId&lt;/span&gt;&lt;span class="ss"&gt;})&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="ss"&gt;[&lt;/span&gt;&lt;span class="nc"&gt;:TARGETS_FIELD&lt;/span&gt;&lt;span class="ss"&gt;]&lt;/span&gt;&lt;span class="o"&gt;-&amp;gt;&lt;/span&gt;&lt;span class="ss"&gt;(&lt;/span&gt;&lt;span class="py"&gt;field:&lt;/span&gt;&lt;span class="n"&gt;SchemaField&lt;/span&gt;&lt;span class="ss"&gt;)&lt;/span&gt;
&lt;span class="k"&gt;MATCH&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="ss"&gt;(&lt;/span&gt;&lt;span class="n"&gt;field&lt;/span&gt;&lt;span class="ss"&gt;)&lt;/span&gt;&lt;span class="o"&gt;&amp;lt;-&lt;/span&gt;&lt;span class="ss"&gt;[&lt;/span&gt;&lt;span class="nc"&gt;:USES_FIELD&lt;/span&gt;&lt;span class="ss"&gt;]&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="ss"&gt;(&lt;/span&gt;&lt;span class="py"&gt;ep:&lt;/span&gt;&lt;span class="n"&gt;Endpoint&lt;/span&gt;&lt;span class="ss"&gt;)&lt;/span&gt;&lt;span class="o"&gt;&amp;lt;-&lt;/span&gt;&lt;span class="ss"&gt;[&lt;/span&gt;&lt;span class="nc"&gt;:CONSUMES&lt;/span&gt;&lt;span class="ss"&gt;]&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="ss"&gt;(&lt;/span&gt;&lt;span class="py"&gt;consumer:&lt;/span&gt;&lt;span class="n"&gt;Service&lt;/span&gt;&lt;span class="ss"&gt;)&lt;/span&gt;
&lt;span class="k"&gt;MATCH&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="ss"&gt;(&lt;/span&gt;&lt;span class="n"&gt;consumer&lt;/span&gt;&lt;span class="ss"&gt;)&lt;/span&gt;&lt;span class="o"&gt;&amp;lt;-&lt;/span&gt;&lt;span class="ss"&gt;[&lt;/span&gt;&lt;span class="nc"&gt;:OWNS&lt;/span&gt;&lt;span class="ss"&gt;]&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="ss"&gt;(&lt;/span&gt;&lt;span class="py"&gt;consumerTeam:&lt;/span&gt;&lt;span class="n"&gt;Team&lt;/span&gt;&lt;span class="ss"&gt;)&lt;/span&gt;
&lt;span class="k"&gt;OPTIONAL&lt;/span&gt; &lt;span class="k"&gt;MATCH&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="ss"&gt;(&lt;/span&gt;&lt;span class="n"&gt;consumerTeam&lt;/span&gt;&lt;span class="ss"&gt;)&lt;/span&gt;&lt;span class="o"&gt;&amp;lt;-&lt;/span&gt;&lt;span class="ss"&gt;[&lt;/span&gt;&lt;span class="nc"&gt;:MEMBER_OF&lt;/span&gt;&lt;span class="ss"&gt;]&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="ss"&gt;(&lt;/span&gt;&lt;span class="py"&gt;lead:&lt;/span&gt;&lt;span class="n"&gt;Person&lt;/span&gt;&lt;span class="ss"&gt;)&lt;/span&gt;
&lt;span class="k"&gt;WHERE&lt;/span&gt; &lt;span class="n"&gt;lead.role&lt;/span&gt; &lt;span class="ow"&gt;CONTAINS&lt;/span&gt; &lt;span class="s1"&gt;'Lead'&lt;/span&gt; &lt;span class="ow"&gt;OR&lt;/span&gt; &lt;span class="n"&gt;lead.role&lt;/span&gt; &lt;span class="ow"&gt;CONTAINS&lt;/span&gt; &lt;span class="s1"&gt;'Staff'&lt;/span&gt;
&lt;span class="k"&gt;RETURN&lt;/span&gt; &lt;span class="k"&gt;DISTINCT&lt;/span&gt;
  &lt;span class="n"&gt;consumer.name&lt;/span&gt; &lt;span class="k"&gt;AS&lt;/span&gt; &lt;span class="n"&gt;serviceName&lt;/span&gt;&lt;span class="ss"&gt;,&lt;/span&gt;
  &lt;span class="n"&gt;consumer.tier&lt;/span&gt; &lt;span class="k"&gt;AS&lt;/span&gt; &lt;span class="n"&gt;serviceTier&lt;/span&gt;&lt;span class="ss"&gt;,&lt;/span&gt;
  &lt;span class="n"&gt;consumerTeam.name&lt;/span&gt; &lt;span class="k"&gt;AS&lt;/span&gt; &lt;span class="n"&gt;owningTeam&lt;/span&gt;&lt;span class="ss"&gt;,&lt;/span&gt;
  &lt;span class="n"&gt;consumerTeam.slack_channel&lt;/span&gt; &lt;span class="k"&gt;AS&lt;/span&gt; &lt;span class="n"&gt;slackChannel&lt;/span&gt;&lt;span class="ss"&gt;,&lt;/span&gt;
  &lt;span class="n"&gt;lead.handle&lt;/span&gt; &lt;span class="k"&gt;AS&lt;/span&gt; &lt;span class="n"&gt;techLeadHandle&lt;/span&gt;&lt;span class="ss"&gt;,&lt;/span&gt;
  &lt;span class="n"&gt;ep.path&lt;/span&gt; &lt;span class="k"&gt;AS&lt;/span&gt; &lt;span class="n"&gt;viaEndpoint&lt;/span&gt;&lt;span class="ss"&gt;;&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;In a single query taking &lt;strong&gt;4ms&lt;/strong&gt;, FalkorDB traverses from the modified schema field through exposed API endpoints to find every consumer service and their corresponding Tech Lead.&lt;/p&gt;

&lt;h3&gt;
  
  
  3. Recalling Cross-Session Episodic Memory
&lt;/h3&gt;

&lt;p&gt;When an ADR is violated, teams don't always reject the PR. Sometimes, a production outage or migration forces a temporary workaround.&lt;/p&gt;

&lt;p&gt;ArchGuard captures these workarounds as &lt;code&gt;:ArchitecturalEpisode&lt;/code&gt; nodes. When a new violation occurs, Agent 3 recalls historical precedents:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight cypher"&gt;&lt;code&gt;&lt;span class="k"&gt;MATCH&lt;/span&gt;&lt;span class="w"&gt; &lt;/span&gt;&lt;span class="ss"&gt;(&lt;/span&gt;&lt;span class="py"&gt;p:&lt;/span&gt;&lt;span class="n"&gt;Proposal&lt;/span&gt; &lt;span class="ss"&gt;{&lt;/span&gt;&lt;span class="py"&gt;id:&lt;/span&gt; &lt;span class="n"&gt;$proposalId&lt;/span&gt;&lt;span class="ss"&gt;})&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="ss"&gt;[&lt;/span&gt;&lt;span class="nc"&gt;:VIOLATES_ADR&lt;/span&gt;&lt;span class="ss"&gt;]&lt;/span&gt;&lt;span class="o"&gt;-&amp;gt;&lt;/span&gt;&lt;span class="ss"&gt;(&lt;/span&gt;&lt;span class="py"&gt;adr:&lt;/span&gt;&lt;span class="n"&gt;ADR&lt;/span&gt;&lt;span class="ss"&gt;)&lt;/span&gt;&lt;span class="o"&gt;&amp;lt;-&lt;/span&gt;&lt;span class="ss"&gt;[&lt;/span&gt;&lt;span class="nc"&gt;:APPLIES_TO&lt;/span&gt;&lt;span class="ss"&gt;]&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="ss"&gt;(&lt;/span&gt;&lt;span class="py"&gt;ep:&lt;/span&gt;&lt;span class="n"&gt;ArchitecturalEpisode&lt;/span&gt;&lt;span class="ss"&gt;)&lt;/span&gt;
&lt;span class="k"&gt;RETURN&lt;/span&gt; 
  &lt;span class="n"&gt;ep.incident_id&lt;/span&gt; &lt;span class="k"&gt;AS&lt;/span&gt; &lt;span class="n"&gt;incidentId&lt;/span&gt;&lt;span class="ss"&gt;,&lt;/span&gt;
  &lt;span class="n"&gt;ep.title&lt;/span&gt; &lt;span class="k"&gt;AS&lt;/span&gt; &lt;span class="n"&gt;title&lt;/span&gt;&lt;span class="ss"&gt;,&lt;/span&gt;
  &lt;span class="n"&gt;ep.workaround_pattern&lt;/span&gt; &lt;span class="k"&gt;AS&lt;/span&gt; &lt;span class="n"&gt;workaroundPattern&lt;/span&gt;&lt;span class="ss"&gt;,&lt;/span&gt;
  &lt;span class="n"&gt;ep.approved_by&lt;/span&gt; &lt;span class="k"&gt;AS&lt;/span&gt; &lt;span class="n"&gt;approvedBy&lt;/span&gt;&lt;span class="ss"&gt;,&lt;/span&gt;
  &lt;span class="n"&gt;ep.resolution_date&lt;/span&gt; &lt;span class="k"&gt;AS&lt;/span&gt; &lt;span class="n"&gt;resolutionDate&lt;/span&gt;&lt;span class="ss"&gt;,&lt;/span&gt;
  &lt;span class="n"&gt;ep.trade_offs&lt;/span&gt; &lt;span class="k"&gt;AS&lt;/span&gt; &lt;span class="n"&gt;tradeOffs&lt;/span&gt;&lt;span class="ss"&gt;;&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;For &lt;code&gt;ADR-001&lt;/code&gt;, FalkorDB immediately recalls &lt;strong&gt;INC-2023-42&lt;/strong&gt;:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;&lt;em&gt;"Legacy Billing System Integer Ingestion Workaround. Approved by @sarah_arch. Workaround: Maintain dual-representation in API gateway by exposing an X-Legacy-User-Id header while keeping the canonical body payload as UUIDv4."&lt;/em&gt;&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;Instead of an AI reviewer hallucinating advice, Agent 3 presents the &lt;strong&gt;exact battle-tested compromise approved by company architects&lt;/strong&gt;.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fipaq8q21hnm01uddvtx4.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fipaq8q21hnm01uddvtx4.png" alt="Episodic Memory Recall UI" width="800" height="450"&gt;&lt;/a&gt;&lt;br&gt;
&lt;em&gt;Figure 2: Memory &amp;amp; Plan tab surfacing historical incident compromise INC-2023-42 with approved workaround steps and trade-offs.&lt;/em&gt;&lt;/p&gt;


&lt;h2&gt;
  
  
  Engineering Lessons &amp;amp; Real Bugs We Fixed
&lt;/h2&gt;

&lt;p&gt;Building on a graph database gave us incredible superpowers, but we hit three real-world engineering hurdles that taught us valuable lessons:&lt;/p&gt;
&lt;h3&gt;
  
  
  1. Deduplicating Blast Radius Paths vs. Distinct Services
&lt;/h3&gt;

&lt;p&gt;During early testing with PR #409, our UI reported: &lt;em&gt;"Breaks 5 downstream consumers"&lt;/em&gt;. But when we opened the microservices list, only 3 services were listed (&lt;code&gt;billing-service&lt;/code&gt;, &lt;code&gt;checkout-service&lt;/code&gt;, and &lt;code&gt;ledger-service&lt;/code&gt;).&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;The Bug:&lt;/strong&gt; &lt;code&gt;billing-service&lt;/code&gt; consumed two different endpoints (&lt;code&gt;/v1/users&lt;/code&gt; and &lt;code&gt;/v1/ledger/entries&lt;/code&gt;). A naive path count treated each traversal path as a separate broken consumer, inflating the count from 3 to 5.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;The Fix:&lt;/strong&gt; We refactored &lt;code&gt;src/engine/graphAuditor.ts&lt;/code&gt; with an in-memory aggregation map:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight typescript"&gt;&lt;code&gt;&lt;span class="c1"&gt;// Deduplicate by distinct serviceName (counting distinct services, not paths)&lt;/span&gt;
&lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="nx"&gt;serviceMap&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="k"&gt;new&lt;/span&gt; &lt;span class="nb"&gt;Map&lt;/span&gt;&lt;span class="o"&gt;&amp;lt;&lt;/span&gt;&lt;span class="kr"&gt;string&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nx"&gt;BlastRadiusFinding&lt;/span&gt;&lt;span class="o"&gt;&amp;gt;&lt;/span&gt;&lt;span class="p"&gt;();&lt;/span&gt;

&lt;span class="k"&gt;for &lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="nx"&gt;r&lt;/span&gt; &lt;span class="k"&gt;of&lt;/span&gt; &lt;span class="nx"&gt;rawRows&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
  &lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="nx"&gt;existing&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nx"&gt;serviceMap&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;get&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nx"&gt;r&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nx"&gt;serviceName&lt;/span&gt;&lt;span class="p"&gt;);&lt;/span&gt;
  &lt;span class="k"&gt;if &lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nx"&gt;existing&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
    &lt;span class="k"&gt;if &lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nx"&gt;r&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nx"&gt;viaEndpoint&lt;/span&gt; &lt;span class="o"&gt;&amp;amp;&amp;amp;&lt;/span&gt; &lt;span class="o"&gt;!&lt;/span&gt;&lt;span class="nx"&gt;existing&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nx"&gt;viaEndpoint&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;includes&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nx"&gt;r&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nx"&gt;viaEndpoint&lt;/span&gt;&lt;span class="p"&gt;))&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
      &lt;span class="nx"&gt;existing&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nx"&gt;viaEndpoint&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="s2"&gt;`&lt;/span&gt;&lt;span class="p"&gt;${&lt;/span&gt;&lt;span class="nx"&gt;existing&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nx"&gt;viaEndpoint&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="s2"&gt;, &lt;/span&gt;&lt;span class="p"&gt;${&lt;/span&gt;&lt;span class="nx"&gt;r&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nx"&gt;viaEndpoint&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="s2"&gt;`&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
    &lt;span class="p"&gt;}&lt;/span&gt;
    &lt;span class="k"&gt;continue&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
  &lt;span class="p"&gt;}&lt;/span&gt;
  &lt;span class="nx"&gt;serviceMap&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;set&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nx"&gt;r&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nx"&gt;serviceName&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="p"&gt;...&lt;/span&gt; &lt;span class="p"&gt;});&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;
&lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="nb"&gt;Array&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="k"&gt;from&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nx"&gt;serviceMap&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;values&lt;/span&gt;&lt;span class="p"&gt;());&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Now, the blast radius accurately reflects &lt;strong&gt;3 distinct Tier-1 microservices&lt;/strong&gt;, while displaying all collapsed endpoint paths in the details card.&lt;/p&gt;

&lt;h3&gt;
  
  
  2. The Latency Split: Real-Time Graph Traversal vs. LLM Time
&lt;/h3&gt;

&lt;p&gt;One of our proudest architectural decisions was displaying the latency split directly in the Telemetry and Cypher Trace panels:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;⚡ FalkorDB Graph Traversal:&lt;/strong&gt; Individual Cypher queries execute in &lt;code&gt;0.8ms – 14ms&lt;/code&gt; (with the full multi-agent 5-query traversal pipeline executing in &lt;code&gt;23ms – 108ms&lt;/code&gt;)&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;⏱️ LLM Synthesis Time:&lt;/strong&gt; &lt;code&gt;~1,200ms – 4,000ms&lt;/code&gt;
&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Seeing this in real-time demonstrates the immense power of in-memory graph databases for AI agents. Graph traversal is virtually instantaneous; LLM token streaming is the bottleneck.&lt;/p&gt;

&lt;p&gt;Furthermore, we added a &lt;strong&gt;zero-violation fast path&lt;/strong&gt;: If a developer fixes their code and changes &lt;code&gt;int64 user_id&lt;/code&gt; back to &lt;code&gt;string user_id&lt;/code&gt;, Agent 2 detects &lt;strong&gt;0 violations and 0 blast radius&lt;/strong&gt;. ArchGuard skips the remediation LLM entirely, turns the proposal node emerald green (&lt;code&gt;:Proposal (Approved)&lt;/code&gt;), and finishes the audit in milliseconds!&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fi94ouk2z7r35gaow5cxt.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fi94ouk2z7r35gaow5cxt.png" alt="Green Glow Approved Proposal" width="800" height="450"&gt;&lt;/a&gt;&lt;br&gt;
&lt;em&gt;Figure 3: Compliant proposals glow green on the pitch-black constellation canvas with zero blast radius.&lt;/em&gt;&lt;/p&gt;
&lt;h3&gt;
  
  
  3. Bulletproof Offline Heuristic Fallback
&lt;/h3&gt;

&lt;p&gt;During hackathon live demos or production network blips, external LLM APIs can hit rate limits or time out. &lt;/p&gt;

&lt;p&gt;To ensure ArchGuard never fails on stage, we built a zero-dependency deterministic heuristic fallback in &lt;code&gt;src/llm/client.ts&lt;/code&gt;. If no API keys are provided or if external calls fail, our AST pattern matcher extracts schema types and generates compliant migration templates in &lt;code&gt;0.0s&lt;/code&gt;. The system is 100% resilient.&lt;/p&gt;


&lt;h2&gt;
  
  
  Live ADR Ingestion &amp;amp; Memory Recording
&lt;/h2&gt;

&lt;p&gt;ArchGuard isn't a static demo—it is a live, evolving Company Brain:&lt;/p&gt;
&lt;h3&gt;
  
  
  1. Live Markdown ADR Ingestion (&lt;code&gt;+ Ingest ADR (.md)&lt;/code&gt;)
&lt;/h3&gt;

&lt;p&gt;Click the &lt;strong&gt;&lt;code&gt;+ Ingest ADR&lt;/code&gt;&lt;/strong&gt; button in the header, paste standard markdown with YAML frontmatter:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight markdown"&gt;&lt;code&gt;&lt;span class="nn"&gt;---&lt;/span&gt;
&lt;span class="na"&gt;code&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="s"&gt;ADR-005&lt;/span&gt;
&lt;span class="na"&gt;title&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="s"&gt;Enforce Idempotency Keys on Payment Ingestion&lt;/span&gt;
&lt;span class="na"&gt;status&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="s"&gt;Accepted&lt;/span&gt;
&lt;span class="na"&gt;governs_target&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="s"&gt;/v1/webhooks/payment&lt;/span&gt;
&lt;span class="na"&gt;target_property&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="s"&gt;idempotency_enforced&lt;/span&gt;
&lt;span class="na"&gt;expected_value&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"&lt;/span&gt;&lt;span class="s"&gt;true"&lt;/span&gt;
&lt;span class="na"&gt;author_handle&lt;/span&gt;&lt;span class="pi"&gt;:&lt;/span&gt; &lt;span class="s"&gt;elena_fin&lt;/span&gt;
&lt;span class="nn"&gt;---&lt;/span&gt;
&lt;span class="gu"&gt;## Context&lt;/span&gt;
Double-charging incidents require mandatory Idempotency-Key headers on all payment webhooks.
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Click &lt;strong&gt;Compile into FalkorDB Brain&lt;/strong&gt;. FalkorDB dynamically parses the frontmatter, merges the &lt;code&gt;:ADR&lt;/code&gt; node, creates the &lt;code&gt;[:GOVERNS]-&amp;gt;(:Endpoint)&lt;/code&gt; edge, and links &lt;code&gt;[:AUTHORED]-&amp;gt;(:Person)&lt;/code&gt; in real time. The node counter ticks up from 28 to 29 instantly.&lt;/p&gt;

&lt;h3&gt;
  
  
  2. Recording New Episodic Memory (&lt;code&gt;+ Record Memory&lt;/code&gt;)
&lt;/h3&gt;

&lt;p&gt;When an audit flags a violation, engineers can click &lt;strong&gt;&lt;code&gt;+ Commit New Workaround to FalkorDB Memory&lt;/code&gt;&lt;/strong&gt;. This records a new &lt;code&gt;:ArchitecturalEpisode&lt;/code&gt; linked via &lt;code&gt;[:APPLIES_TO]&lt;/code&gt; to the violated ADR. The next engineer who tests a similar change will immediately see that workaround in their audit results.&lt;/p&gt;




&lt;h2&gt;
  
  
  Try It Yourself
&lt;/h2&gt;

&lt;p&gt;ArchGuard is fully open source. You can clone the repo and run the full stack locally in under two minutes:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;&lt;span class="c"&gt;# 1. Clone repository&lt;/span&gt;
git clone https://github.com/SKfaizan-786/ArchGuard.git
&lt;span class="nb"&gt;cd &lt;/span&gt;ArchGuard

&lt;span class="c"&gt;# 2. Launch FalkorDB via Docker&lt;/span&gt;
docker compose up &lt;span class="nt"&gt;-d&lt;/span&gt;

&lt;span class="c"&gt;# 3. Install dependencies &amp;amp; launch full-stack&lt;/span&gt;
npm &lt;span class="nb"&gt;install
&lt;/span&gt;npm &lt;span class="nt"&gt;--prefix&lt;/span&gt; client &lt;span class="nb"&gt;install
&lt;/span&gt;npm run dev:all
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;ArchGuard UI:&lt;/strong&gt; &lt;a href="http://localhost:5173" rel="noopener noreferrer"&gt;http://localhost:5173&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;FalkorDB Web Browser:&lt;/strong&gt; &lt;a href="http://localhost:3000" rel="noopener noreferrer"&gt;http://localhost:3000&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Backend API:&lt;/strong&gt; &lt;a href="http://localhost:4000/api/health" rel="noopener noreferrer"&gt;http://localhost:4000/api/health&lt;/a&gt;
&lt;/li&gt;
&lt;/ul&gt;

&lt;h3&gt;
  
  
  Project Links
&lt;/h3&gt;

&lt;ul&gt;
&lt;li&gt;💻 &lt;strong&gt;GitHub Repository:&lt;/strong&gt; &lt;a href="https://github.com/SKfaizan-786/ArchGuard" rel="noopener noreferrer"&gt;https://github.com/SKfaizan-786/ArchGuard&lt;/a&gt;
&lt;/li&gt;
&lt;li&gt;🎥 &lt;strong&gt;Live Demo Video:&lt;/strong&gt; &lt;a href="https://youtu.be/qKTABIh2Djs" rel="noopener noreferrer"&gt;https://youtu.be/qKTABIh2Djs&lt;/a&gt;
&lt;/li&gt;
&lt;/ul&gt;




&lt;h2&gt;
  
  
  Final Thoughts: The Future of Agentic Governance
&lt;/h2&gt;

&lt;p&gt;AI coding agents are already generating pull requests faster than human tech leads can review them. If we rely on probabilistic LLMs to review probabilistic LLMs, architectural drift is guaranteed.&lt;/p&gt;

&lt;p&gt;By anchoring AI agents to a &lt;strong&gt;deterministic FalkorDB knowledge graph&lt;/strong&gt;, we get the best of both worlds:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;The Graph&lt;/strong&gt; provides mathematical certainty, instant multi-hop reachability, and persistent episodic memory across sessions.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;The Agents&lt;/strong&gt; provide human-friendly natural language synthesis, intelligent diff generation, and contextual remediation.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Linters check your syntax. ArchGuard protects your architecture.&lt;/p&gt;




&lt;p&gt;&lt;em&gt;Built with ❤️ by SK Faizanuddin for the &lt;strong&gt;WeMakeDevs x FalkorDB "Graph&lt;/strong&gt; &lt;strong&gt;Hacks" Hackathon&lt;/strong&gt; (Company Brain &amp;amp; Episodic Memory Track).&lt;/em&gt;&lt;/p&gt;

</description>
      <category>wemakedevs</category>
      <category>falkordb</category>
      <category>ai</category>
      <category>architecture</category>
    </item>
    <item>
      <title>I Built a Real-Time AI Interview Coach Using Vision Agents — Here's How</title>
      <dc:creator>SK Faizanuddin</dc:creator>
      <pubDate>Sun, 01 Mar 2026 18:00:53 +0000</pubDate>
      <link>https://dev.to/skfaizan786/i-built-a-real-time-ai-interview-coach-using-vision-agents-heres-how-50k9</link>
      <guid>https://dev.to/skfaizan786/i-built-a-real-time-ai-interview-coach-using-vision-agents-heres-how-50k9</guid>
      <description>&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.amazonaws.com%2Fuploads%2Farticles%2F5qjfxdh2dbplkpknll0a.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.amazonaws.com%2Fuploads%2Farticles%2F5qjfxdh2dbplkpknll0a.png" alt="AceView landing page with AI interview coach interface and video call layout" width="800" height="424"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;The Problem That Started It All&lt;/strong&gt;&lt;br&gt;
You've been preparing for a job interview for weeks. Your answers are sharp. Your research is solid. But when the actual interview happens — you slouch, you break eye contact, you say "um" seventeen times, and you speak so fast nobody can follow you. The frustrating part? You knew better. You just didn't have anyone watching.&lt;br&gt;
That's the gap AceView fills.&lt;br&gt;
&lt;strong&gt;AceView&lt;/strong&gt; is a real-time AI interview coach that joins your video call, watches you with computer vision, listens to you with speech recognition, and gives you live feedback on six performance axes — every single second.&lt;br&gt;
Not after the interview. During it.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;The Vision: What We Set Out to Build&lt;/strong&gt;&lt;br&gt;
The dream was simple to describe and hard to build:&lt;/p&gt;

&lt;p&gt;&lt;em&gt;"An AI that watches you practice an interview and coaches you the way a great mentor would — not by grading you at the end, but by whispering in your ear the whole time.&lt;/em&gt;"&lt;/p&gt;

&lt;p&gt;Six things to track, in real time, for every second of every session:&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Signal&lt;/strong&gt;        &lt;strong&gt;What We Measure&lt;/strong&gt;&lt;br&gt;
👁️ Eye Contact Are you looking at the camera?&lt;br&gt;
🧍 Posture    Are your shoulders straight?&lt;br&gt;
💬 Filler Words   How many "um", "like", "you know"?&lt;br&gt;
🎙️ Speech Pace Words per minute vs ideal 130 WPM&lt;br&gt;
🤖 AI Nudges  Smart coaching tips fired when you slip&lt;br&gt;
📊 Report Card    Gemini-generated grade + actionable feedback&lt;br&gt;
The result: a platform where you can do a full mock interview, get coached live by an ElevenLabs-voiced AI, and walk away with a PDF report card — all in one session.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;The Execution: How Vision Agents Made This Possible&lt;/strong&gt;&lt;br&gt;
The core of AceView runs on the &lt;strong&gt;Vision Agents SDK by GetStream&lt;/strong&gt;. Before building this, I was genuinely unsure if it was possible to do pose detection and speech analysis simultaneously on a live video call without building custom WebRTC infrastructure from scratch. Vision Agents solved the hardest parts.&lt;br&gt;
Here's how the system is wired:&lt;/p&gt;

&lt;p&gt;Frontend (Next.js)&lt;br&gt;
    │&lt;br&gt;
    │ WebRTC (Stream Video SDK)&lt;br&gt;
    ▼&lt;br&gt;
Backend (FastAPI)&lt;br&gt;
    │&lt;br&gt;
    ├── Vision Agents Agent (joins call as participant)&lt;br&gt;
    │       ├── YOLOPoseProcessor → posture + eye contact&lt;br&gt;
    │       ├── Deepgram STT → speech → WPM + fillers&lt;br&gt;
    │       ├── ElevenLabs TTS → AI coach voice&lt;br&gt;
    │       └── Gemini 2.0 Flash → interview questions + report&lt;br&gt;
    │&lt;br&gt;
    └── Custom events → frontend metrics panel&lt;/p&gt;

&lt;p&gt;The agent joins the video call as a second participant. It sees the user's video stream, hears their audio, and sends coaching data back to the frontend — all through the same WebRTC connection.&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.amazonaws.com%2Fuploads%2Farticles%2Fzk33ow9cvrtd0qvpab2j.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.amazonaws.com%2Fuploads%2Farticles%2Fzk33ow9cvrtd0qvpab2j.png" alt="AceView Live Session — HIGH Confidence _The confidence ring glows green when posture and eye contact are both strong. Metrics update every second on the right panel._" width="800" height="422"&gt;&lt;/a&gt;&lt;br&gt;
&lt;strong&gt;AceView Live Session&lt;/strong&gt; — HIGH Confidence &lt;em&gt;The confidence ring glows green when posture and eye contact are both strong. Metrics update every second on the right panel.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Setting Up the Agent&lt;/strong&gt;&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight python"&gt;&lt;code&gt;&lt;span class="kn"&gt;from&lt;/span&gt; &lt;span class="n"&gt;vision_agents.core&lt;/span&gt; &lt;span class="kn"&gt;import&lt;/span&gt; &lt;span class="n"&gt;Agent&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;AgentLauncher&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;Runner&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;User&lt;/span&gt;
&lt;span class="kn"&gt;from&lt;/span&gt; &lt;span class="n"&gt;vision_agents.plugins&lt;/span&gt; &lt;span class="kn"&gt;import&lt;/span&gt; &lt;span class="n"&gt;deepgram&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;elevenlabs&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;openai&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;getstream&lt;/span&gt;
&lt;span class="kn"&gt;from&lt;/span&gt; &lt;span class="n"&gt;agents.vision_processor&lt;/span&gt; &lt;span class="kn"&gt;import&lt;/span&gt; &lt;span class="n"&gt;AceViewVisionProcessor&lt;/span&gt;
&lt;span class="k"&gt;async&lt;/span&gt; &lt;span class="k"&gt;def&lt;/span&gt; &lt;span class="nf"&gt;create_agent&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="o"&gt;**&lt;/span&gt;&lt;span class="n"&gt;kwargs&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;-&amp;gt;&lt;/span&gt; &lt;span class="n"&gt;Agent&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
    &lt;span class="n"&gt;llm&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;openai&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nc"&gt;ChatCompletionsLLM&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;
        &lt;span class="n"&gt;api_key&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="n"&gt;os&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;getenv&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;OPENROUTER_API_KEY&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;),&lt;/span&gt;
        &lt;span class="n"&gt;base_url&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;https://openrouter.ai/api/v1&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
        &lt;span class="n"&gt;model&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;google/gemini-2.0-flash-001&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;
    &lt;span class="p"&gt;)&lt;/span&gt;
    &lt;span class="n"&gt;agent&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nc"&gt;Agent&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;
        &lt;span class="n"&gt;edge&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="n"&gt;getstream&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nc"&gt;Edge&lt;/span&gt;&lt;span class="p"&gt;(),&lt;/span&gt;  &lt;span class="c1"&gt;# Low-latency edge network
&lt;/span&gt;        &lt;span class="n"&gt;agent_user&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="nc"&gt;User&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;name&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;AceView AI Coach&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nb"&gt;id&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;aceview_agent&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;),&lt;/span&gt;
        &lt;span class="n"&gt;instructions&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="n"&gt;SYSTEM_PROMPT&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
        &lt;span class="n"&gt;processors&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;
            &lt;span class="nc"&gt;AceViewVisionProcessor&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;model_path&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="n"&gt;MODEL_PATH&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;fps&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="mi"&gt;1&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;conf_threshold&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="mf"&gt;0.25&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
        &lt;span class="p"&gt;],&lt;/span&gt;
        &lt;span class="n"&gt;llm&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="n"&gt;llm&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
        &lt;span class="n"&gt;tts&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="n"&gt;elevenlabs&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nc"&gt;TTS&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;model_id&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;eleven_flash_v2_5&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;),&lt;/span&gt;
        &lt;span class="n"&gt;stt&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="n"&gt;deepgram&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nc"&gt;STT&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;eager_turn_detection&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="bp"&gt;True&lt;/span&gt;&lt;span class="p"&gt;),&lt;/span&gt;
        &lt;span class="n"&gt;streaming_tts&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="bp"&gt;True&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;  &lt;span class="c1"&gt;# Drastically cuts response latency
&lt;/span&gt;    &lt;span class="p"&gt;)&lt;/span&gt;
    &lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="n"&gt;agent&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Notice the &lt;strong&gt;processors&lt;/strong&gt; list — that's our custom vision processor injected directly into the agent pipeline. Vision Agents calls it every frame with the YOLO keypoints. We don't write any WebRTC or frame-extraction code. The SDK handles all of that.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;The Code: Real-Time Vision Analysis&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;Building a Custom Vision Processor&lt;br&gt;
The SDK's &lt;strong&gt;YOLOPoseProcessor&lt;/strong&gt; is the key class we extend. You inherit from it, override &lt;strong&gt;process()&lt;/strong&gt;, and you get every YOLO pose frame automatically.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight python"&gt;&lt;code&gt;&lt;span class="kn"&gt;from&lt;/span&gt; &lt;span class="n"&gt;vision_agents.plugins.ultralytics&lt;/span&gt; &lt;span class="kn"&gt;import&lt;/span&gt; &lt;span class="n"&gt;YOLOPoseProcessor&lt;/span&gt;
&lt;span class="kn"&gt;from&lt;/span&gt; &lt;span class="n"&gt;vision_agents.core&lt;/span&gt; &lt;span class="kn"&gt;import&lt;/span&gt; &lt;span class="n"&gt;Agent&lt;/span&gt;
&lt;span class="kn"&gt;import&lt;/span&gt; &lt;span class="n"&gt;numpy&lt;/span&gt; &lt;span class="k"&gt;as&lt;/span&gt; &lt;span class="n"&gt;np&lt;/span&gt;
&lt;span class="k"&gt;class&lt;/span&gt; &lt;span class="nc"&gt;AceViewVisionProcessor&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;YOLOPoseProcessor&lt;/span&gt;&lt;span class="p"&gt;):&lt;/span&gt;
    &lt;span class="sh"&gt;"""&lt;/span&gt;&lt;span class="s"&gt;
    Custom YOLO processor that calculates real-time interview coaching metrics
    and broadcasts them to the frontend via WebRTC custom events.
    &lt;/span&gt;&lt;span class="sh"&gt;"""&lt;/span&gt;
    &lt;span class="n"&gt;name&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;aceview_vision&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;
    &lt;span class="k"&gt;def&lt;/span&gt; &lt;span class="nf"&gt;__init__&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;self&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="o"&gt;*&lt;/span&gt;&lt;span class="n"&gt;args&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="o"&gt;**&lt;/span&gt;&lt;span class="n"&gt;kwargs&lt;/span&gt;&lt;span class="p"&gt;):&lt;/span&gt;
        &lt;span class="n"&gt;kwargs&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;setdefault&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;fps&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="mi"&gt;1&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
        &lt;span class="nf"&gt;super&lt;/span&gt;&lt;span class="p"&gt;().&lt;/span&gt;&lt;span class="nf"&gt;__init__&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="o"&gt;*&lt;/span&gt;&lt;span class="n"&gt;args&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="o"&gt;**&lt;/span&gt;&lt;span class="n"&gt;kwargs&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
        &lt;span class="n"&gt;self&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;agent&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;Optional&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="n"&gt;Agent&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="bp"&gt;None&lt;/span&gt;
        &lt;span class="n"&gt;self&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;_last_nudge_times&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;Dict&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="nb"&gt;str&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nb"&gt;float&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="p"&gt;{}&lt;/span&gt;
    &lt;span class="k"&gt;def&lt;/span&gt; &lt;span class="nf"&gt;attach_agent&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;self&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;agent&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;Agent&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;-&amp;gt;&lt;/span&gt; &lt;span class="bp"&gt;None&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
        &lt;span class="sh"&gt;"""&lt;/span&gt;&lt;span class="s"&gt;Called automatically by the SDK when the agent starts.&lt;/span&gt;&lt;span class="sh"&gt;"""&lt;/span&gt;
        &lt;span class="n"&gt;self&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;agent&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;agent&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The SDK automatically calls &lt;strong&gt;attach_agent()&lt;/strong&gt; — so by the time the session starts, our processor has a reference to the agent and can send custom events directly to the frontend.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Posture Scoring with YOLO Keypoints&lt;/strong&gt;&lt;br&gt;
YOLO pose detection gives us 17 keypoints per person — every joint and facial landmark with confidence scores. We use shoulder alignment, torso height, and head position to compute a posture score:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight python"&gt;&lt;code&gt;&lt;span class="c1"&gt;# COCO keypoint indices
&lt;/span&gt;&lt;span class="n"&gt;LEFT_SHOULDER&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;RIGHT_SHOULDER&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="mi"&gt;5&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="mi"&gt;6&lt;/span&gt;
&lt;span class="n"&gt;LEFT_HIP&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;RIGHT_HIP&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="mi"&gt;11&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="mi"&gt;12&lt;/span&gt;
&lt;span class="n"&gt;NOSE&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="mi"&gt;0&lt;/span&gt;
&lt;span class="n"&gt;CONF_THRESH&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="mf"&gt;0.3&lt;/span&gt;
&lt;span class="k"&gt;def&lt;/span&gt; &lt;span class="nf"&gt;_calculate_posture_score&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;self&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;kpts&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;np&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;ndarray&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;-&amp;gt;&lt;/span&gt; &lt;span class="nb"&gt;int&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
    &lt;span class="sh"&gt;"""&lt;/span&gt;&lt;span class="s"&gt;Score 0-100 from shoulder level, torso height, and head position.&lt;/span&gt;&lt;span class="sh"&gt;"""&lt;/span&gt;
    &lt;span class="n"&gt;scores&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="p"&gt;[]&lt;/span&gt;
    &lt;span class="n"&gt;l_sh&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;r_sh&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;kpts&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="n"&gt;LEFT_SHOULDER&lt;/span&gt;&lt;span class="p"&gt;],&lt;/span&gt; &lt;span class="n"&gt;kpts&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="n"&gt;RIGHT_SHOULDER&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;
    &lt;span class="n"&gt;l_hip&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;r_hip&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;kpts&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="n"&gt;LEFT_HIP&lt;/span&gt;&lt;span class="p"&gt;],&lt;/span&gt; &lt;span class="n"&gt;kpts&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="n"&gt;RIGHT_HIP&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;
    &lt;span class="n"&gt;nose&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;kpts&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="n"&gt;NOSE&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;
    &lt;span class="c1"&gt;# 1. Shoulder tilt (tilted shoulders = slouching)
&lt;/span&gt;    &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="n"&gt;l_sh&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="mi"&gt;2&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt; &lt;span class="o"&gt;&amp;gt;&lt;/span&gt; &lt;span class="n"&gt;CONF_THRESH&lt;/span&gt; &lt;span class="ow"&gt;and&lt;/span&gt; &lt;span class="n"&gt;r_sh&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="mi"&gt;2&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt; &lt;span class="o"&gt;&amp;gt;&lt;/span&gt; &lt;span class="n"&gt;CONF_THRESH&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
        &lt;span class="n"&gt;shoulder_width&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nf"&gt;abs&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;l_sh&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="mi"&gt;0&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt; &lt;span class="o"&gt;-&lt;/span&gt; &lt;span class="n"&gt;r_sh&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="mi"&gt;0&lt;/span&gt;&lt;span class="p"&gt;])&lt;/span&gt;
        &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="n"&gt;shoulder_width&lt;/span&gt; &lt;span class="o"&gt;&amp;gt;&lt;/span&gt; &lt;span class="mi"&gt;10&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
            &lt;span class="n"&gt;tilt&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nf"&gt;abs&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;l_sh&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="mi"&gt;1&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt; &lt;span class="o"&gt;-&lt;/span&gt; &lt;span class="n"&gt;r_sh&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="mi"&gt;1&lt;/span&gt;&lt;span class="p"&gt;])&lt;/span&gt; &lt;span class="o"&gt;/&lt;/span&gt; &lt;span class="n"&gt;shoulder_width&lt;/span&gt;
            &lt;span class="n"&gt;scores&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;append&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nf"&gt;max&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="mi"&gt;10&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nf"&gt;int&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="mi"&gt;100&lt;/span&gt; &lt;span class="o"&gt;-&lt;/span&gt; &lt;span class="n"&gt;tilt&lt;/span&gt; &lt;span class="o"&gt;*&lt;/span&gt; &lt;span class="mi"&gt;300&lt;/span&gt;&lt;span class="p"&gt;)))&lt;/span&gt;
    &lt;span class="c1"&gt;# 2. Torso upright (hips below shoulders = sitting straight)
&lt;/span&gt;    &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="nf"&gt;all&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;kpts&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="n"&gt;i&lt;/span&gt;&lt;span class="p"&gt;][&lt;/span&gt;&lt;span class="mi"&gt;2&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt; &lt;span class="o"&gt;&amp;gt;&lt;/span&gt; &lt;span class="n"&gt;CONF_THRESH&lt;/span&gt; &lt;span class="k"&gt;for&lt;/span&gt; &lt;span class="n"&gt;i&lt;/span&gt; &lt;span class="ow"&gt;in&lt;/span&gt; &lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="n"&gt;LEFT_SHOULDER&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;RIGHT_SHOULDER&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;LEFT_HIP&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;RIGHT_HIP&lt;/span&gt;&lt;span class="p"&gt;]):&lt;/span&gt;
        &lt;span class="n"&gt;avg_sh_y&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;l_sh&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="mi"&gt;1&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt; &lt;span class="o"&gt;+&lt;/span&gt; &lt;span class="n"&gt;r_sh&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="mi"&gt;1&lt;/span&gt;&lt;span class="p"&gt;])&lt;/span&gt; &lt;span class="o"&gt;/&lt;/span&gt; &lt;span class="mi"&gt;2&lt;/span&gt;
        &lt;span class="n"&gt;avg_hip_y&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;l_hip&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="mi"&gt;1&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt; &lt;span class="o"&gt;+&lt;/span&gt; &lt;span class="n"&gt;r_hip&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="mi"&gt;1&lt;/span&gt;&lt;span class="p"&gt;])&lt;/span&gt; &lt;span class="o"&gt;/&lt;/span&gt; &lt;span class="mi"&gt;2&lt;/span&gt;
        &lt;span class="n"&gt;torso_h&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;avg_hip_y&lt;/span&gt; &lt;span class="o"&gt;-&lt;/span&gt; &lt;span class="n"&gt;avg_sh_y&lt;/span&gt;
        &lt;span class="n"&gt;scores&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;append&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="mi"&gt;90&lt;/span&gt; &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="n"&gt;torso_h&lt;/span&gt; &lt;span class="o"&gt;&amp;gt;&lt;/span&gt; &lt;span class="mi"&gt;30&lt;/span&gt; &lt;span class="k"&gt;else&lt;/span&gt; &lt;span class="mi"&gt;65&lt;/span&gt; &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="n"&gt;torso_h&lt;/span&gt; &lt;span class="o"&gt;&amp;gt;&lt;/span&gt; &lt;span class="mi"&gt;0&lt;/span&gt; &lt;span class="k"&gt;else&lt;/span&gt; &lt;span class="mi"&gt;35&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
    &lt;span class="c1"&gt;# 3. Head upright (nose above shoulder line)
&lt;/span&gt;    &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="n"&gt;nose&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="mi"&gt;2&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt; &lt;span class="o"&gt;&amp;gt;&lt;/span&gt; &lt;span class="n"&gt;CONF_THRESH&lt;/span&gt; &lt;span class="ow"&gt;and&lt;/span&gt; &lt;span class="n"&gt;l_sh&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="mi"&gt;2&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt; &lt;span class="o"&gt;&amp;gt;&lt;/span&gt; &lt;span class="n"&gt;CONF_THRESH&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
        &lt;span class="n"&gt;avg_sh_y&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;l_sh&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="mi"&gt;1&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt; &lt;span class="o"&gt;+&lt;/span&gt; &lt;span class="n"&gt;r_sh&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="mi"&gt;1&lt;/span&gt;&lt;span class="p"&gt;])&lt;/span&gt; &lt;span class="o"&gt;/&lt;/span&gt; &lt;span class="mi"&gt;2&lt;/span&gt;
        &lt;span class="n"&gt;scores&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;append&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="mi"&gt;95&lt;/span&gt; &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="n"&gt;nose&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="mi"&gt;1&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt; &lt;span class="o"&gt;&amp;lt;&lt;/span&gt; &lt;span class="n"&gt;avg_sh_y&lt;/span&gt; &lt;span class="k"&gt;else&lt;/span&gt; &lt;span class="mi"&gt;40&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
    &lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="nf"&gt;min&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="mi"&gt;100&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nf"&gt;int&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nf"&gt;sum&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;scores&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;/&lt;/span&gt; &lt;span class="nf"&gt;len&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;scores&lt;/span&gt;&lt;span class="p"&gt;)))&lt;/span&gt; &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="n"&gt;scores&lt;/span&gt; &lt;span class="k"&gt;else&lt;/span&gt; &lt;span class="mi"&gt;0&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;strong&gt;Eye Contact via Ear Asymmetry (The Key Insight)&lt;/strong&gt;&lt;br&gt;
YOLO can't track eyeball direction. But there's a clever proxy: &lt;strong&gt;ear visibility&lt;/strong&gt;. &lt;br&gt;
When you look left, your right ear becomes more visible. When you look right, your left ear pops into view. Perfect forward gaze = both ears visible with similar confidence. Head turned = major asymmetry.&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight python"&gt;&lt;code&gt;&lt;span class="k"&gt;def&lt;/span&gt; &lt;span class="nf"&gt;_calculate_eye_contact&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;self&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;kpts&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;np&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;ndarray&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;-&amp;gt;&lt;/span&gt; &lt;span class="n"&gt;Tuple&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="nb"&gt;int&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nb"&gt;bool&lt;/span&gt;&lt;span class="p"&gt;]:&lt;/span&gt;
    &lt;span class="sh"&gt;"""&lt;/span&gt;&lt;span class="s"&gt;
    Strategy: YOLO can&lt;/span&gt;&lt;span class="sh"&gt;'&lt;/span&gt;&lt;span class="s"&gt;t track eyeballs, but it CAN detect head rotation
    via which ears are visible. Ear asymmetry = head-rotation signal.
    &lt;/span&gt;&lt;span class="sh"&gt;"""&lt;/span&gt;
    &lt;span class="n"&gt;l_ear_conf&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;kpts&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="n"&gt;LEFT_EAR&lt;/span&gt;&lt;span class="p"&gt;][&lt;/span&gt;&lt;span class="mi"&gt;2&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt; &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="n"&gt;kpts&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="n"&gt;LEFT_EAR&lt;/span&gt;&lt;span class="p"&gt;][&lt;/span&gt;&lt;span class="mi"&gt;2&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt; &lt;span class="o"&gt;&amp;gt;&lt;/span&gt; &lt;span class="n"&gt;CONF_THRESH&lt;/span&gt; &lt;span class="k"&gt;else&lt;/span&gt; &lt;span class="mf"&gt;0.0&lt;/span&gt;
    &lt;span class="n"&gt;r_ear_conf&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;kpts&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="n"&gt;RIGHT_EAR&lt;/span&gt;&lt;span class="p"&gt;][&lt;/span&gt;&lt;span class="mi"&gt;2&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt; &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="n"&gt;kpts&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="n"&gt;RIGHT_EAR&lt;/span&gt;&lt;span class="p"&gt;][&lt;/span&gt;&lt;span class="mi"&gt;2&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt; &lt;span class="o"&gt;&amp;gt;&lt;/span&gt; &lt;span class="n"&gt;CONF_THRESH&lt;/span&gt; &lt;span class="k"&gt;else&lt;/span&gt; &lt;span class="mf"&gt;0.0&lt;/span&gt;
    &lt;span class="n"&gt;ear_total&lt;/span&gt;  &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;l_ear_conf&lt;/span&gt; &lt;span class="o"&gt;+&lt;/span&gt; &lt;span class="n"&gt;r_ear_conf&lt;/span&gt;
    &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="n"&gt;ear_total&lt;/span&gt; &lt;span class="o"&gt;&amp;gt;&lt;/span&gt; &lt;span class="mf"&gt;0.05&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
        &lt;span class="c1"&gt;# 0 = symmetric (facing camera), 1 = only one ear visible (turned head)
&lt;/span&gt;        &lt;span class="n"&gt;ear_asymmetry&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nf"&gt;abs&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;l_ear_conf&lt;/span&gt; &lt;span class="o"&gt;-&lt;/span&gt; &lt;span class="n"&gt;r_ear_conf&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;/&lt;/span&gt; &lt;span class="n"&gt;ear_total&lt;/span&gt;
        &lt;span class="n"&gt;ear_penalty&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nf"&gt;int&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;ear_asymmetry&lt;/span&gt; &lt;span class="o"&gt;*&lt;/span&gt; &lt;span class="mi"&gt;80&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;  &lt;span class="c1"&gt;# up to 80-point drop
&lt;/span&gt;    &lt;span class="k"&gt;else&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
        &lt;span class="n"&gt;ear_penalty&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="mi"&gt;0&lt;/span&gt;
    &lt;span class="c1"&gt;# Refine with nose horizontal offset
&lt;/span&gt;    &lt;span class="n"&gt;nose_penalty&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="mi"&gt;0&lt;/span&gt;
    &lt;span class="n"&gt;l_eye&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;r_eye&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;nose&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;kpts&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="n"&gt;LEFT_EYE&lt;/span&gt;&lt;span class="p"&gt;],&lt;/span&gt; &lt;span class="n"&gt;kpts&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="n"&gt;RIGHT_EYE&lt;/span&gt;&lt;span class="p"&gt;],&lt;/span&gt; &lt;span class="n"&gt;kpts&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="n"&gt;NOSE&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;
    &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="nf"&gt;all&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;kpts&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="n"&gt;i&lt;/span&gt;&lt;span class="p"&gt;][&lt;/span&gt;&lt;span class="mi"&gt;2&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt; &lt;span class="o"&gt;&amp;gt;&lt;/span&gt; &lt;span class="n"&gt;CONF_THRESH&lt;/span&gt; &lt;span class="k"&gt;for&lt;/span&gt; &lt;span class="n"&gt;i&lt;/span&gt; &lt;span class="ow"&gt;in&lt;/span&gt; &lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="n"&gt;LEFT_EYE&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;RIGHT_EYE&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;NOSE&lt;/span&gt;&lt;span class="p"&gt;]):&lt;/span&gt;
        &lt;span class="n"&gt;eye_cx&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;l_eye&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="mi"&gt;0&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt; &lt;span class="o"&gt;+&lt;/span&gt; &lt;span class="n"&gt;r_eye&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="mi"&gt;0&lt;/span&gt;&lt;span class="p"&gt;])&lt;/span&gt; &lt;span class="o"&gt;/&lt;/span&gt; &lt;span class="mi"&gt;2&lt;/span&gt;
        &lt;span class="n"&gt;eye_w&lt;/span&gt;  &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nf"&gt;abs&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;l_eye&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="mi"&gt;0&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt; &lt;span class="o"&gt;-&lt;/span&gt; &lt;span class="n"&gt;r_eye&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="mi"&gt;0&lt;/span&gt;&lt;span class="p"&gt;])&lt;/span&gt;
        &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="n"&gt;eye_w&lt;/span&gt; &lt;span class="o"&gt;&amp;gt;&lt;/span&gt; &lt;span class="mi"&gt;5&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
            &lt;span class="n"&gt;offset&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nf"&gt;abs&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;nose&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="mi"&gt;0&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt; &lt;span class="o"&gt;-&lt;/span&gt; &lt;span class="n"&gt;eye_cx&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;/&lt;/span&gt; &lt;span class="n"&gt;eye_w&lt;/span&gt;
            &lt;span class="n"&gt;nose_penalty&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nf"&gt;int&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;offset&lt;/span&gt; &lt;span class="o"&gt;*&lt;/span&gt; &lt;span class="mi"&gt;100&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
    &lt;span class="n"&gt;score&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nf"&gt;max&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="mi"&gt;5&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="mi"&gt;100&lt;/span&gt; &lt;span class="o"&gt;-&lt;/span&gt; &lt;span class="nf"&gt;min&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="mi"&gt;95&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;ear_penalty&lt;/span&gt; &lt;span class="o"&gt;+&lt;/span&gt; &lt;span class="n"&gt;nose_penalty&lt;/span&gt;&lt;span class="p"&gt;))&lt;/span&gt;
    &lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="n"&gt;score&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="bp"&gt;True&lt;/span&gt;

&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;This approach made the eye contact score far more sensitive than the naive "nose offset" method I'd tried first. Looking away from the camera now drops the score to 25–40 range almost instantly.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Real-Time Filler Word Detection + WPM Tracking&lt;/strong&gt;&lt;br&gt;
Speech processing runs through Deepgram's streaming STT. We subscribe to two events — partial transcripts (live preview) and final transcripts (committed sentences). On every final transcript, we count filler words and compute the real words-per-minute:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight python"&gt;&lt;code&gt;&lt;span class="n"&gt;FILLER_WORDS&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
    &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;umm&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;um&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;hmm&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;hm&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;mm&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;mhm&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;uh&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;like&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;you know&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;basically&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;actually&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;literally&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;right&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;so&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;
&lt;span class="c1"&gt;# Tracks real speaking time for WPM
&lt;/span&gt;&lt;span class="n"&gt;_pace&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;words&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="mi"&gt;0&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;speaking_secs&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="mf"&gt;0.0&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;turn_start&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="bp"&gt;None&lt;/span&gt;&lt;span class="p"&gt;}&lt;/span&gt;
&lt;span class="k"&gt;async&lt;/span&gt; &lt;span class="k"&gt;def&lt;/span&gt; &lt;span class="nf"&gt;on_final_transcript&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;event&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;STTTranscriptEvent&lt;/span&gt;&lt;span class="p"&gt;):&lt;/span&gt;
    &lt;span class="n"&gt;text&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;event&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;text&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;strip&lt;/span&gt;&lt;span class="p"&gt;()&lt;/span&gt;
    &lt;span class="n"&gt;normalized&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;text&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;lower&lt;/span&gt;&lt;span class="p"&gt;()&lt;/span&gt;
    &lt;span class="c1"&gt;# Count fillers (multi-word phrases first, then single words)
&lt;/span&gt;    &lt;span class="n"&gt;filler_count&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nf"&gt;len&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;re&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;findall&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="sa"&gt;r&lt;/span&gt;&lt;span class="sh"&gt;'&lt;/span&gt;&lt;span class="s"&gt;\byou know\b&lt;/span&gt;&lt;span class="sh"&gt;'&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;normalized&lt;/span&gt;&lt;span class="p"&gt;))&lt;/span&gt;
    &lt;span class="n"&gt;words&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;re&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;split&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="sa"&gt;r&lt;/span&gt;&lt;span class="sh"&gt;'&lt;/span&gt;&lt;span class="s"&gt;\W+&lt;/span&gt;&lt;span class="sh"&gt;'&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;normalized&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
    &lt;span class="n"&gt;filler_count&lt;/span&gt; &lt;span class="o"&gt;+=&lt;/span&gt; &lt;span class="nf"&gt;sum&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="mi"&gt;1&lt;/span&gt; &lt;span class="k"&gt;for&lt;/span&gt; &lt;span class="n"&gt;w&lt;/span&gt; &lt;span class="ow"&gt;in&lt;/span&gt; &lt;span class="n"&gt;words&lt;/span&gt; &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="n"&gt;w&lt;/span&gt; &lt;span class="ow"&gt;in&lt;/span&gt; &lt;span class="n"&gt;FILLER_WORDS&lt;/span&gt; &lt;span class="o"&gt;-&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;you know&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;})&lt;/span&gt;
    &lt;span class="c1"&gt;# Accumulate real speaking time
&lt;/span&gt;    &lt;span class="n"&gt;_pace&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;words&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt; &lt;span class="o"&gt;+=&lt;/span&gt; &lt;span class="nf"&gt;len&lt;/span&gt;&lt;span class="p"&gt;([&lt;/span&gt;&lt;span class="n"&gt;w&lt;/span&gt; &lt;span class="k"&gt;for&lt;/span&gt; &lt;span class="n"&gt;w&lt;/span&gt; &lt;span class="ow"&gt;in&lt;/span&gt; &lt;span class="n"&gt;words&lt;/span&gt; &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="n"&gt;w&lt;/span&gt;&lt;span class="p"&gt;])&lt;/span&gt;
    &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="n"&gt;_pace&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;turn_start&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt; &lt;span class="ow"&gt;is&lt;/span&gt; &lt;span class="ow"&gt;not&lt;/span&gt; &lt;span class="bp"&gt;None&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
        &lt;span class="n"&gt;_pace&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;speaking_secs&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt; &lt;span class="o"&gt;+=&lt;/span&gt; &lt;span class="n"&gt;time&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;monotonic&lt;/span&gt;&lt;span class="p"&gt;()&lt;/span&gt; &lt;span class="o"&gt;-&lt;/span&gt; &lt;span class="n"&gt;_pace&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;turn_start&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;
        &lt;span class="n"&gt;_pace&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;turn_start&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="bp"&gt;None&lt;/span&gt;
    &lt;span class="c1"&gt;# 130 WPM = perfect score of 100
&lt;/span&gt;    &lt;span class="n"&gt;pace_score&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="mi"&gt;82&lt;/span&gt;  &lt;span class="c1"&gt;# default until we have enough data
&lt;/span&gt;    &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="n"&gt;_pace&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;speaking_secs&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt; &lt;span class="o"&gt;&amp;gt;=&lt;/span&gt; &lt;span class="mf"&gt;3.0&lt;/span&gt; &lt;span class="ow"&gt;and&lt;/span&gt; &lt;span class="n"&gt;_pace&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;words&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt; &lt;span class="o"&gt;&amp;gt;=&lt;/span&gt; &lt;span class="mi"&gt;5&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
        &lt;span class="n"&gt;wpm&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;_pace&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;words&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt; &lt;span class="o"&gt;/&lt;/span&gt; &lt;span class="n"&gt;_pace&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;speaking_secs&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;])&lt;/span&gt; &lt;span class="o"&gt;*&lt;/span&gt; &lt;span class="mi"&gt;60&lt;/span&gt;
        &lt;span class="n"&gt;pace_score&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nf"&gt;max&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="mi"&gt;0&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nf"&gt;min&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="mi"&gt;100&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
            &lt;span class="nf"&gt;int&lt;/span&gt;&lt;span class="p"&gt;((&lt;/span&gt;&lt;span class="n"&gt;wpm&lt;/span&gt; &lt;span class="o"&gt;-&lt;/span&gt; &lt;span class="mi"&gt;50&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;/&lt;/span&gt; &lt;span class="mi"&gt;80&lt;/span&gt; &lt;span class="o"&gt;*&lt;/span&gt; &lt;span class="mi"&gt;100&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="n"&gt;wpm&lt;/span&gt; &lt;span class="o"&gt;&amp;lt;=&lt;/span&gt; &lt;span class="mi"&gt;130&lt;/span&gt;
            &lt;span class="k"&gt;else&lt;/span&gt; &lt;span class="nf"&gt;int&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="mi"&gt;100&lt;/span&gt; &lt;span class="o"&gt;-&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;wpm&lt;/span&gt; &lt;span class="o"&gt;-&lt;/span&gt; &lt;span class="mi"&gt;130&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;/&lt;/span&gt; &lt;span class="mi"&gt;80&lt;/span&gt; &lt;span class="o"&gt;*&lt;/span&gt; &lt;span class="mi"&gt;100&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
        &lt;span class="p"&gt;))&lt;/span&gt;
    &lt;span class="k"&gt;await&lt;/span&gt; &lt;span class="n"&gt;agent&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;send_custom_event&lt;/span&gt;&lt;span class="p"&gt;({&lt;/span&gt;
        &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;type&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;transcript&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
        &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;text&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;text&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
        &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;filler_count&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;filler_count&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
        &lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;pace_score&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="n"&gt;pace_score&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
    &lt;span class="p"&gt;})&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.amazonaws.com%2Fuploads%2Farticles%2Fyql1tgio6kastvnb292z.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.amazonaws.com%2Fuploads%2Farticles%2Fyql1tgio6kastvnb292z.png" alt="Filler Words Highlighted in Live Transcript _Every filler word is highlighted in real time as the user speaks. The counter on the right updates live._" width="800" height="420"&gt;&lt;/a&gt;&lt;br&gt;
Filler Words Highlighted in Live Transcript &lt;em&gt;Every filler word is highlighted in real time as the user speaks. The counter on the right updates live.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;agent.send_custom_event()&lt;/strong&gt; broadcasts this directly to the frontend over WebRTC. No REST API, no polling. The metrics panel updates in real time.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;AI Nudges — Invisible Mid-Session Coaching&lt;/strong&gt;&lt;br&gt;
When a metric crosses a threshold, we fire a coaching nudge — a small pop-up that appears for a few seconds and disappears. Each nudge type has its own 10-second cooldown, so multiple issues can fire simultaneously without overwhelming the user:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight python"&gt;&lt;code&gt;&lt;span class="n"&gt;NUDGE_COOLDOWN&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="mf"&gt;10.0&lt;/span&gt;  &lt;span class="c1"&gt;# seconds
&lt;/span&gt;&lt;span class="k"&gt;def&lt;/span&gt; &lt;span class="nf"&gt;_should_nudge&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;self&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;key&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="nb"&gt;str&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;-&amp;gt;&lt;/span&gt; &lt;span class="nb"&gt;bool&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
    &lt;span class="n"&gt;now&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;time&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;time&lt;/span&gt;&lt;span class="p"&gt;()&lt;/span&gt;
    &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="n"&gt;now&lt;/span&gt; &lt;span class="o"&gt;-&lt;/span&gt; &lt;span class="n"&gt;self&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;_last_nudge_times&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;get&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;key&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="mf"&gt;0.0&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;&amp;gt;=&lt;/span&gt; &lt;span class="n"&gt;NUDGE_COOLDOWN&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
        &lt;span class="n"&gt;self&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;_last_nudge_times&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="n"&gt;key&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;now&lt;/span&gt;
        &lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="bp"&gt;True&lt;/span&gt;
    &lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="bp"&gt;False&lt;/span&gt;
&lt;span class="k"&gt;async&lt;/span&gt; &lt;span class="k"&gt;def&lt;/span&gt; &lt;span class="nf"&gt;_send_nudge_if_needed&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;self&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;posture&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="nb"&gt;int&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;eye_contact&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="nb"&gt;int&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="n"&gt;face_visible&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="nb"&gt;bool&lt;/span&gt;&lt;span class="p"&gt;):&lt;/span&gt;
    &lt;span class="c1"&gt;# Independent if-checks — all three can fire simultaneously
&lt;/span&gt;    &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="ow"&gt;not&lt;/span&gt; &lt;span class="n"&gt;face_visible&lt;/span&gt; &lt;span class="ow"&gt;and&lt;/span&gt; &lt;span class="n"&gt;self&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;_should_nudge&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;face&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;):&lt;/span&gt;
        &lt;span class="k"&gt;await&lt;/span&gt; &lt;span class="n"&gt;self&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;_safe_nudge&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;Make sure your face and body are visible on camera&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
    &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="n"&gt;posture&lt;/span&gt; &lt;span class="o"&gt;&amp;lt;&lt;/span&gt; &lt;span class="mi"&gt;70&lt;/span&gt; &lt;span class="ow"&gt;and&lt;/span&gt; &lt;span class="n"&gt;self&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;_should_nudge&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;posture&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;):&lt;/span&gt;
        &lt;span class="k"&gt;await&lt;/span&gt; &lt;span class="n"&gt;self&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;_safe_nudge&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;Sit up straight and square your shoulders to the camera&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
    &lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="n"&gt;eye_contact&lt;/span&gt; &lt;span class="o"&gt;&amp;lt;&lt;/span&gt; &lt;span class="mi"&gt;65&lt;/span&gt; &lt;span class="ow"&gt;and&lt;/span&gt; &lt;span class="n"&gt;self&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;_should_nudge&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;eye&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;):&lt;/span&gt;
        &lt;span class="k"&gt;await&lt;/span&gt; &lt;span class="n"&gt;self&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;_safe_nudge&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="s"&gt;Look directly at your camera — eye contact matters in interviews&lt;/span&gt;&lt;span class="sh"&gt;"&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.amazonaws.com%2Fuploads%2Farticles%2F5q1keqfbf9fsok41u5gl.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.amazonaws.com%2Fuploads%2Farticles%2F5q1keqfbf9fsok41u5gl.png" alt="AI Nudge Firing — Bad Eye Contact _A nudge appears mid-session when eye contact drops below threshold. It fades out automatically — no interruption to the interview flow._" width="800" height="424"&gt;&lt;/a&gt;&lt;br&gt;
&lt;strong&gt;AI Nudge Firing&lt;/strong&gt; — Bad Eye Contact &lt;em&gt;A nudge appears mid-session when eye contact drops below threshold. It fades out automatically — no interruption to the interview flow.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Session Averaging: Fairness by Design&lt;/strong&gt;&lt;br&gt;
One design decision I'm proud of: the report card is based on &lt;strong&gt;the average of the entire session&lt;/strong&gt;, not the score at the moment you click "End Session." Early on, the report card reflected only the last 5 seconds of the call — meaning if you adjusted your posture right before ending, you'd get an inflated score. We fixed this with a running accumulator on the frontend:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight typescript"&gt;&lt;code&gt;&lt;span class="c1"&gt;// Zustand store — accumulates every frame for true session averaging&lt;/span&gt;
&lt;span class="kr"&gt;interface&lt;/span&gt; &lt;span class="nx"&gt;SessionAccumulator&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
  &lt;span class="nl"&gt;postureSum&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="kr"&gt;number&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt; &lt;span class="nl"&gt;eyeSum&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="kr"&gt;number&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt; &lt;span class="nl"&gt;paceSum&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="kr"&gt;number&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt; &lt;span class="nl"&gt;count&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="kr"&gt;number&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;
&lt;span class="nl"&gt;endSession&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="k"&gt;async &lt;/span&gt;&lt;span class="p"&gt;()&lt;/span&gt; &lt;span class="o"&gt;=&amp;gt;&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
  &lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="nx"&gt;_acc&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nx"&gt;metrics&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nf"&gt;get&lt;/span&gt;&lt;span class="p"&gt;();&lt;/span&gt;
  &lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="nx"&gt;avgPosture&lt;/span&gt;   &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nx"&gt;_acc&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nx"&gt;count&lt;/span&gt; &lt;span class="o"&gt;&amp;gt;&lt;/span&gt; &lt;span class="mi"&gt;0&lt;/span&gt; &lt;span class="p"&gt;?&lt;/span&gt; &lt;span class="nb"&gt;Math&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;round&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nx"&gt;_acc&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nx"&gt;postureSum&lt;/span&gt; &lt;span class="o"&gt;/&lt;/span&gt; &lt;span class="nx"&gt;_acc&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nx"&gt;count&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="nx"&gt;metrics&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nx"&gt;postureScore&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
  &lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="nx"&gt;avgEye&lt;/span&gt;       &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nx"&gt;_acc&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nx"&gt;count&lt;/span&gt; &lt;span class="o"&gt;&amp;gt;&lt;/span&gt; &lt;span class="mi"&gt;0&lt;/span&gt; &lt;span class="p"&gt;?&lt;/span&gt; &lt;span class="nb"&gt;Math&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;round&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nx"&gt;_acc&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nx"&gt;eyeSum&lt;/span&gt;    &lt;span class="o"&gt;/&lt;/span&gt; &lt;span class="nx"&gt;_acc&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nx"&gt;count&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="nx"&gt;metrics&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nx"&gt;eyeContactScore&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
  &lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="nx"&gt;avgPace&lt;/span&gt;      &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nx"&gt;_acc&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nx"&gt;count&lt;/span&gt; &lt;span class="o"&gt;&amp;gt;&lt;/span&gt; &lt;span class="mi"&gt;0&lt;/span&gt; &lt;span class="p"&gt;?&lt;/span&gt; &lt;span class="nb"&gt;Math&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;round&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nx"&gt;_acc&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nx"&gt;paceSum&lt;/span&gt;   &lt;span class="o"&gt;/&lt;/span&gt; &lt;span class="nx"&gt;_acc&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nx"&gt;count&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="nx"&gt;metrics&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nx"&gt;speechPaceScore&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
  &lt;span class="c1"&gt;// Update displayed metrics to show session averages&lt;/span&gt;
  &lt;span class="nf"&gt;set&lt;/span&gt;&lt;span class="p"&gt;({&lt;/span&gt; &lt;span class="na"&gt;isSessionActive&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="kc"&gt;false&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="na"&gt;metrics&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="p"&gt;...&lt;/span&gt;&lt;span class="nx"&gt;metrics&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="na"&gt;postureScore&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="nx"&gt;avgPosture&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="p"&gt;...&lt;/span&gt; &lt;span class="p"&gt;}});&lt;/span&gt;
  &lt;span class="c1"&gt;// Send averages to Gemini for honest report card&lt;/span&gt;
  &lt;span class="nf"&gt;get&lt;/span&gt;&lt;span class="p"&gt;().&lt;/span&gt;&lt;span class="nf"&gt;fetchSummary&lt;/span&gt;&lt;span class="p"&gt;({&lt;/span&gt; &lt;span class="na"&gt;posture&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="nx"&gt;avgPosture&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="na"&gt;eye&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="nx"&gt;avgEye&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="na"&gt;pace&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="nx"&gt;avgPace&lt;/span&gt; &lt;span class="p"&gt;});&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The result: if you had terrible eye contact for 8 minutes and then looked at the camera for the last 30 seconds, your eye contact score will honestly reflect your session performance.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;The Report Card: Honest AI Feedback&lt;/strong&gt;&lt;br&gt;
After the session, averages are sent to Gemini 2.0 Flash via OpenRouter. The prompt includes strict rules to prevent contradictory feedback:&lt;/p&gt;

&lt;p&gt;Metrics ≥ 75 → listed as &lt;strong&gt;strengths&lt;/strong&gt; only&lt;br&gt;
Metrics &amp;lt; 75 → listed as &lt;strong&gt;areas to improve&lt;/strong&gt;&lt;br&gt;
If all scores are low → strengths focus on effort, not weak metrics&lt;br&gt;
Grading: A ≥ 85, B ≥ 70, C ≥ 55, D &amp;lt; 55&lt;br&gt;
No more "Grade D — great posture!" contradictions.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;What We Learned&lt;/strong&gt;&lt;br&gt;
&lt;strong&gt;On Vision Agents SDK&lt;/strong&gt;: The SDK's YOLOPoseProcessor is genuinely powerful. The ability to extend it with a custom class and get frames automatically — without writing a single line of WebRTC video capture code — saved at least a week of work. The agent.send_custom_event() method is elegant: any JSON you send appears as a custom event on the frontend call object.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;On real-time ML&lt;/strong&gt;: Running YOLO at 1 FPS (not 3) was a critical decision. At 3 FPS, the audio queue started starving — ElevenLabs would stutter or drop audio entirely. At 1 FPS, everything runs smoothly. Always profile your ML inference loops in an audio-first pipeline.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;On eye contact detection&lt;/strong&gt;: The "nose offset" approach I tried first was too weak. Ear asymmetry is a much more reliable signal for head rotation because it reflects a physical landmark (the 3D orientation of the head), not a 2D projection artifact.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Final Result&lt;/strong&gt;&lt;br&gt;
✅ 6 real-time coaching signals&lt;br&gt;
✅ ElevenLabs-voiced AI interviewer conducting the call&lt;br&gt;
✅ Live transcript with filler word highlights&lt;br&gt;
✅ Confidence ring animation (green → red, faster pulse = lower confidence)&lt;br&gt;
✅ Mid-session AI nudges for posture, eye contact, and filler words&lt;br&gt;
✅ Session-averaged Gemini report card with honest grading&lt;br&gt;
✅ One-click PDF download&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.amazonaws.com%2Fuploads%2Farticles%2Ffev0227kpobqyphlgli9.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.amazonaws.com%2Fuploads%2Farticles%2Ffev0227kpobqyphlgli9.png" alt="AI Report Card — Grade &amp;amp; Feedback _Gemini generates an honest A–D graded report card. Strengths only appear for metrics ≥ 75 — no fake praise._" width="800" height="399"&gt;&lt;/a&gt;&lt;br&gt;
&lt;strong&gt;AI Report Card&lt;/strong&gt; — Grade &amp;amp; Feedback &lt;em&gt;Gemini generates an honest A–D graded report card. Strengths only appear for metrics ≥ 75 — no fake praise.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;Try It&lt;br&gt;
🔗 &lt;strong&gt;GitHub:&lt;/strong&gt; &lt;a href="https://github.com/SKfaizan-786/aceview" rel="noopener noreferrer"&gt;https://github.com/SKfaizan-786/aceview&lt;/a&gt;&lt;br&gt;
🌐 &lt;strong&gt;Live Demo:&lt;/strong&gt; &lt;a href="https://aceview-murex.vercel.app" rel="noopener noreferrer"&gt;https://aceview-murex.vercel.app&lt;/a&gt;&lt;br&gt;
🎬 &lt;strong&gt;Demo Video:&lt;/strong&gt; &lt;a href="https://youtu.be/PQ_nZ8KGVYQ" rel="noopener noreferrer"&gt;https://youtu.be/PQ_nZ8KGVYQ&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Built with: &lt;strong&gt;Vision Agents · YOLOv11 · Deepgram · ElevenLabs · Gemini 2.0 Flash · Stream Video · Next.js · FastAPI&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;&lt;em&gt;Submitted for Vision Possible Hackathon 2025 by WeMakeDevs — #BuildInPublic #VisionAgents #AIHackathon&lt;/em&gt;&lt;/p&gt;

</description>
      <category>hackathon</category>
      <category>visionagents</category>
      <category>wemakedevs</category>
      <category>ai</category>
    </item>
  </channel>
</rss>
